Nvidia’s SoL-Pi system optimizes the control harness between language models and their coding environments to reduce token usage by up to 49 percent with minimal performance degradation. The research involved testing 152 different approaches across more than 3,000 runs to develop the system, though gains were smaller on other benchmarks. (The Decoder)
Liquid AI released LFM2.5-VL-3B-DSpark, a 279.5M-parameter draft model enabling speculative decoding for its LFM2.5-VL-3B vision-language model. The draft model achieves up to 3.13x faster decoding on Apple M5 Max and 2.66x on H100 hardware with identical output under greedy decoding, with support shipping in llama.cpp and other inference frameworks. (MarkTechPost) Perplexity published a post-training study showing its Computer agent reduced tool-call failures from 2.24 percent to 1.77 percent by training on real user sessions paired with hint-guided self-distillation, demonstrating improvements through learning from mistakes in live deployment. (MarkTechPost)